Pharmacokinetics / Pharmacodynamics

PK/PD Modeling for Individualized Dosing

Patient-specific parameter estimation from sparse clinical data for precision medicine

Introduction

Pharmacokinetics (PK) describes what the body does to a drug—absorption, distribution, and elimination—while pharmacodynamics (PD) describes what the drug does to the body—efficacy and toxicity. Optimal dosing requires patient-specific parameter estimation to account for inter-individual variability in clearance, volume of distribution, and drug sensitivity.

The challenge: clinical data are sparse (3–5 blood samples over days/weeks), irregular (missed appointments), and noisy (assay variability ±10–20%). How do we estimate patient-specific PK/PD parameters to guide dose adjustments, predict exposure, and personalize therapy?

      Why this matters: Individualized dosing improves efficacy (target attainment rate increases 30–50%), reduces toxicity (adverse events decrease 20–40%), and enables therapeutic drug monitoring (TDM) for narrow-window drugs like immunosuppressants, chemotherapy, and antibiotics—improving patient outcomes and reducing healthcare costs.
    

Pharmacological Science

Drug disposition follows ADME principles (Absorption, Distribution, Metabolism, Excretion):

Absorption: Oral drugs enter systemic circulation via gut absorption (first-order kinetics with lag time)
Distribution: Drug distributes between plasma (central compartment) and tissues (peripheral compartments) based on perfusion and partition coefficients
Clearance: Hepatic metabolism (CYP enzymes) and renal excretion remove drug at rate CL·C (first-order elimination)
Exposure-response: Plasma concentration C drives pharmacological effect via Emax models (sigmoidal dose-response)
Biomarker dynamics: Efficacy (tumor size, inflammation markers) and toxicity (QTc prolongation, neutropenia) evolve over hours to weeks

Compartmental Model: Two-Compartment PK with Indirect-Response PD

We use a mechanistic two-compartment model coupled to an indirect-response PD model:

States: $$\begin{aligned} A_g &= \text{drug amount in gut (absorption depot) [mg]} \\ A_c &= \text{drug amount in central compartment (plasma) [mg]} \\ A_p &= \text{drug amount in peripheral compartment (tissues) [mg]} \\ E &= \text{pharmacodynamic effect (biomarker) [units]} \end{aligned}$$ ODEs (PK): $$\begin{aligned} \frac{dA_g}{dt} &= -k_a \cdot A_g \quad \text{[absorption]} \\ \frac{dA_c}{dt} &= k_a \cdot A_g - \frac{CL}{V_c} \cdot A_c - Q \cdot \left(\frac{A_c}{V_c} - \frac{A_p}{V_p}\right) \quad \text{[distribution + elimination]} \\ \frac{dA_p}{dt} &= Q \cdot \left(\frac{A_c}{V_c} - \frac{A_p}{V_p}\right) \quad \text{[peripheral equilibration]} \\[0.5em] C &= \frac{A_c}{V_c} \quad \text{[plasma concentration, mg/L]} \end{aligned}$$ ODE (PD - indirect response model): $$\frac{dE}{dt} = k_{\text{in}} \cdot \left(1 - \frac{I_{\max} \cdot C}{IC_{50} + C}\right) - k_{\text{out}} \cdot E$$ Where: $$\begin{aligned} k_{\text{in}} &= \text{baseline production rate of biomarker} \\ k_{\text{out}} &= \text{elimination rate of biomarker } (E_0 = k_{\text{in}}/k_{\text{out}} \text{ at baseline)} \\ I_{\max} &= \text{maximum inhibition } (0 < I_{\max} \leq 1) \\ IC_{50} &= \text{concentration for 50\% inhibition} \end{aligned}$$ Alternative PD models: $$\begin{aligned} \text{Emax (stimulation): } & \frac{dE}{dt} = k_{\text{in}} \cdot \left(1 + \frac{E_{\max} \cdot C}{EC_{50}+C}\right) - k_{\text{out}} \cdot E \\ \text{Tumor growth: } & \frac{d\text{Tumor}}{dt} = \lambda \cdot \text{Tumor} \cdot \left(1 - \frac{\text{Tumor}}{T_{\max}}\right) - k_{\text{kill}} \cdot C \cdot \text{Tumor} \\ \text{Toxicity: } & \text{Exponential damage accumulation with recovery delay} \end{aligned}$$

Unknown Parameters (10 typical)

Parameter	Description	Units	Inter-Patient CV
ka	Absorption rate constant	h⁻¹	30–60%
CL	Clearance	L/h	40–80% (age, renal function)
Vc	Central volume of distribution	L	20–40%
Vp, Q	Peripheral volume, inter-compartment flow	L, L/h	30–50%
F	Bioavailability (oral dosing)	—	20–50%
IC50	Drug potency	mg/L	50–100%
Imax	Maximum effect	—	20–40%
kin, kout	Biomarker production/elimination	units/h, h⁻¹	40–70%
E0	Baseline biomarker level	units	30–60%

Parameter Estimation Challenge

Given clinical data from a patient:

Dosing record: Time and amount of each dose (oral/IV) [precise]
Sparse PK samples: Plasma concentration C at 3–5 time points (trough, peak, mid-dose) [±10–15% assay error, HPLC/LC-MS]
Efficacy biomarker: Tumor size (RECIST), inflammation marker (CRP), viral load [weekly or biweekly measurements, ±10–30% variability]
Toxicity marker: QTc interval, liver enzymes (ALT, AST), neutrophil count [weekly labs, variable compliance]
Covariates: Weight, age, creatinine clearance, genotype (CYP2C19 polymorphism) [used to inform priors]

The inverse problem uses nonlinear mixed-effects modeling (population PK/PD):

Hierarchical Bayesian framework: Patient-level (individual $i$): $$\begin{aligned} \theta_i &= [CL_i, V_{c,i}, k_{a,i}, IC_{50,i}, \ldots] \quad \text{[individual parameters]} \\ \log(\theta_i) &\sim \mathcal{N}(\log(\theta_{\text{pop}}), \Omega) \quad \text{[log-normal distribution]} \\[0.5em] \theta_{\text{pop}} &= \text{population mean parameters} \\ \Omega &= \text{inter-patient covariance matrix} \end{aligned}$$ Observations: $$\begin{aligned} C_{ij} &= \frac{A_{c,i}(t_j)}{V_{c,i}} + \epsilon_{ij} \quad \text{[PK measurement } j \text{ for patient } i\text{]} \\ \epsilon_{ij} &\sim \mathcal{N}(0, \sigma^2_{\text{PK}}) \quad \text{[proportional error model: } \sigma = 0.15 \cdot C\text{]} \\[0.5em] E_{ij} &= E(t_j; \theta_i) + \eta_{ij} \quad \text{[PD biomarker]} \\ \eta_{ij} &\sim \mathcal{N}(0, \sigma^2_{\text{PD}}) \end{aligned}$$ Maximum a posteriori (MAP) estimation: $$\begin{aligned} \hat{\theta}_i &= \arg\max \left[ \log p(C_i, E_i \mid \theta_i) + \log p(\theta_i \mid \theta_{\text{pop}}, \Omega) \right] \\ &= \arg\min \left[ \sum_j \frac{(C_{ij} - \hat{C}_{ij}(\theta_i))^2}{\sigma^2_{\text{PK}}} + \sum_j \frac{(E_{ij} - \hat{E}_{ij}(\theta_i))^2}{\sigma^2_{\text{PD}}} + (\theta_i - \theta_{\text{pop}})^T \Omega^{-1} (\theta_i - \theta_{\text{pop}}) \right] \end{aligned}$$

      Clinical interpretation: The prior (θpop, Ω from population study) regularizes individual estimates when data are sparse. After 1–2 samples, we get a rough estimate; after 4–5 samples, individual θᵢ converges within ±20% of true value (sufficient for dose adjustment).
    

Virtual Patient Simulation

Generate synthetic clinical trial data:

Define population distribution:
- CL ~ LogNormal(μ=5 L/h, σ=0.5) adjusted for renal function: CL = CLpop · (CrCL/100)^0.75
- Vc ~ LogNormal(μ=40 L, σ=0.3) scaled by weight: Vc = Vcpop · (WT/70)^0.7
- IC50 ~ LogNormal(μ=2 mg/L, σ=0.6) with genetic covariate (CYP2C19 poor metabolizers: IC50 ↑ 50%)
Sample 100 virtual patients: Draw θᵢ from population distributions
Dosing protocol: Standard regimen (e.g., 400 mg BID oral for 28 days)
Simulate ODEs: Forward integration with θᵢ to generate "true" C(t), E(t)
Sparse sampling: Extract PK at t = [2h, 12h, 24h, 168h] post-dose; PD weekly
Add noise: Proportional error (±15% for PK, ±20% for PD)
Missed visits: Randomly drop 10–20% of samples (realistic adherence)

Estimation Workflow

1. Population Analysis (NONMEM, Monolix)

First-stage: fit population parameters from Phase I/II data (50–200 patients)

Estimate θpop, Ω using FOCE (first-order conditional estimation) or SAEM (stochastic approximation EM)
Covariate screening: test renal function, weight, age, genotype for CL/Vc
Validate with VPC (visual predictive checks), goodness-of-fit plots

2. Individual Bayesian Estimation (TDM)

For new patient with 2–5 concentrations:

$$\begin{aligned} &\text{1. Initialize: } \hat{\theta}_i = \theta_{\text{pop}} \quad \text{[use population mean as prior]} \\ &\text{2. Optimize: minimize objective function via L-BFGS-B} \\ &\qquad \Rightarrow \hat{\theta}_i = [CL_i, V_{c,i}, k_{a,i}, IC_{50,i}, \ldots] \\ &\text{3. Predict: simulate } C(t), E(t) \text{ for next 24–168 hours} \\ &\text{4. Dose recommendation:} \\ &\qquad \text{- Target: } AUC_{0-24} = 100 \text{ mg·h/L (efficacy threshold)} \\ &\qquad \text{- Constraint: } C_{\max} < 15 \text{ mg/L (toxicity limit)} \\ &\qquad \text{- Adjust dose: } D_{\text{new}} = D_{\text{old}} \cdot \frac{AUC_{\text{target}}}{AUC_{\text{predicted}}} \\ &\text{5. Update after next measurement (iterative refinement)} \end{aligned}$$

3. Extended Kalman Filter (Real-Time Adaptive Dosing)

For intensive care or oncology infusions:

Augmented state: x = [Ag, Ac, Ap, E, CL, Vc]
Process noise: allows CL to drift due to organ dysfunction
Update EKF whenever new C or E measurement arrives
Adjust infusion rate in real-time based on predicted exposure

4. Markov Chain Monte Carlo (MCMC)

For full posterior distribution and uncertainty quantification:

Sample p(θᵢ | Cᵢ, Eᵢ, θpop, Ω) using Stan or PyMC
Obtain 95% credible intervals for dose recommendations
Useful for dose-finding in special populations (pediatrics, hepatic impairment)

Identifiability from Sparse Sampling

Key insights from sensitivity analysis:

      Minimum data requirements:
      Clearance CL: Requires ≥2 samples after steady-state (trough + one mid-dose)
Volume Vc: Needs peak sample within 0.5–2h post-dose (distribution phase)
Absorption ka: Oral dosing requires ≥3 samples in absorption phase (0.5h, 1h, 2h); IV dosing: ka not identifiable (not needed)
IC50: Needs PD measurements spanning 0.2×IC50 to 5×IC50 (wide concentration range)
Biomarker kinetics (kin, kout): Identifiable only if measurements span ≥2 half-lives of E (e.g., if t½ = 48h, need 4+ days of data)

    

Optimal sampling: D-optimal design maximizes det(Fisher Information Matrix). For two-compartment model, optimal times are typically: [0.5h, 2h, 6h, 24h] post-dose. Rich sampling (8–10 points) in Phase I reduces to sparse sampling (3–4 points) in TDM.

Validation Strategy

Validation Test	Method	Acceptance Criterion
Parameter recovery	100 virtual patients, estimate θᵢ from sparse samples	Median bias < 15%, 90% CI coverage ≥ 85%
Exposure prediction	Predict AUC from 2 samples, compare to true AUC	MPE < 20%, R² > 0.80
Dose recommendation accuracy	Simulate recommended dose, check if target AUC achieved	80% of patients within 80–125% of target
Biomarker forecasting	Predict E at week 4 from weeks 1–2 data	RMSE < 30% of baseline range
Clinical trial simulation	Compare fixed-dose vs. TDM-guided in 500 virtual patients	TDM improves target attainment by ≥ 30%

Clinical Impact

Patient-specific PK/PD modeling delivers measurable improvements:

Therapeutic drug monitoring (TDM): Individualized dosing for vancomycin, tacrolimus, and methotrexate reduces toxicity-related hospitalizations by 25–40% (cost savings: $5K–20K per prevented event)
Oncology dose optimization: Model-informed precision dosing (MIPD) for targeted therapies increases progression-free survival by 2–4 months in metastatic cancer (HR 0.7–0.8)
Pediatric dosing: Allometric scaling + Bayesian estimation enable first-cycle dose accuracy in children (reduces trial-and-error from 3–4 cycles to 1–2)
Renal/hepatic impairment: PK models with organ-function covariates prevent under/overdosing in special populations (FDA guidance encourages model-based approaches)
Drug-drug interaction prediction: Mechanistic PK models predict CYP inhibition effects, guiding combination therapy dosing

      Case study: A hospital implemented Bayesian TDM software for vancomycin across 500 patients. Target trough attainment (10–15 mg/L) increased from 45% to 78%, acute kidney injury rates dropped from 18% to 11%, and pharmacist time per patient decreased from 25 to 10 minutes. Annual savings: $1.2M in reduced toxicity and improved cure rates.
    

Try ProcessLM

ProcessLM automates PK/PD workflows: describe your drug (oral/IV, one/two-compartment), upload clinical data (dosing + concentrations + biomarkers), and it fits population parameters, generates individual estimates, and recommends optimal doses—no NONMEM scripting or population PK expertise required.

Request Early Access